Welcome to BigMother Labs!
BigMother Labs (BML) is a Cambridge-based non-profit Artificial General Intelligence (AGI) research lab founded by Aaron Turner. Aaron is currently BML's sole researcher.
BML's primary focus is the ASI Endgame Problem (AEP): “How can we ensure that the ASI Endgame that actually transpires is as close as possible (and ideally identical) to the ideal ASI Endgame (from the perspective of the human species as a whole)?” This is a combined AGI Alignment and AGI Governance problem — accordingly, we seek to address both aspects.
Aaron Turner: mini bio, research agenda, LinkedIn
Paper: "An Implementation-Neutral solution to the Normative ASI Superalignment Problem"
The latest version of this paper is available as a preprint here.
The paper proposes a specific top-level objective (TTQ) for a superintelligent agent, together with a set of conditions (NAP) on the agent’s precursor, and argues that the two together ensure that the agent’s deployed behaviour is aligned with human values and preferences — thereby reducing the problem of building an aligned superintelligence to the problem of building a precursor that satisfies the NAP conditions.
Along the way, the paper includes:
-
definitions (for the purposes of the paper) of intelligence, AI, AGI, ASI
-
definitions of existentially-critical, safety-critical, and non-critical systems
-
definitions of alignment, superalignment, validation, and supervalidation
-
definitions of extrinsic vs intrinsic intelligence
-
definitions of the three scalable dimensions of intelligence (inventiveness, information, and resources)
-
a detailed consideration of alignment, including:
-
safety and liveness
-
utility functions
-
human preferences and values
-
preference elicitation
-
preference aggregation
-
preference realisation
-
maximal-alignment, indifferent-alignment, and maximal-misalignment
-
consciousness and its implications for alignment
-
The paper poses two challenges to readers:
-
can you find a scenario where the NAP conditions hold and the agent is nonetheless misaligned?
-
can you build an NAP-compliant precursor?
Supplemental material: paper evolution
Further papers are planned and will be announced on this website.
BigMother Labs is a non-profit AGI lab based in Cambridge (UK).
We currently rely entirely on purely philanthropic donations.
Please donate whatever you can (via the above button) to help us continue our work.
Thank you!
Testimonials
Professor Leslie Smith
(University of Stirling, Associate Editor of the Journal of Artificial General Intelligence)
“Much of what passes for AI is simply neural networks, basically the same ideas as in the 1980’s but upgraded with various bells and whistles like new activation functions, summarising levels, and feedback. Even the most sophisticated systems (like the GPT systems) lack novel ideas. Aaron’s work aims to push beyond this, to set an agenda for actual intelligent systems (so called artificial general intelligence, AGI) that considers more than pattern recognition and synthetic language construction. This is quite different from what is pursued by companies, and most computing departments. The work is important, and may underlie the next steps forward in artificial intelligence.”
Professor Steve Young CBE FRS
(University of Cambridge, Chair of Information Engineering)
“The TTQ paper is certainly a tour de force. Aaron sets out a carefully argued process for producing an AGI in as safe a manner as possible. I hope that people read it and at minimum use it as a check list of things to consider.”